A communication line engineering surveying and designing method and system using unmanned aerial vehicle surveying

Abstract

The application discloses a communication line engineering surveying and designing method and system using unmanned aerial vehicle surveying, and the method comprises the following steps: S1, acquiring unmanned aerial vehicle image data collected by an unmanned aerial vehicle on a surveying area; S2, performing preset control point piercing and aerial triangulation encryption processing on the unmanned aerial vehicle image data to generate a digital elevation model (DEM) and a digital orthographic image (DOM), which are used as surface model data of the surveying area to reconstruct a three-dimensional terrain; S3, extracting key points in the three-dimensional terrain based on an optimized U-Net model; S4, selecting a set of path key points in the three-dimensional terrain based on an ant colony routing selection algorithm, and then performing arrangement and combination on multiple paths between the key points to generate multiple primary target routes; and S5, performing comprehensive evaluation on the primary target routes by using a comprehensive evaluation method based on analytic hierarchy process to obtain an optimal target route. The application has the advantages of high efficiency, high automation degree and low cost.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the field of unmanned aerial vehicle surveying and mapping technology, and particularly relates to a communication line engineering survey and design method and system using unmanned aerial vehicle surveying and mapping. BACKGROUND

[0002] Unmanned aerial vehicle surveying and mapping is a new direction of applying unmanned aerial vehicle technology to the field of surveying and mapping remote sensing, and is a high integration of new surveying and mapping remote sensing technology, aviation platform technology, information technology and sensor technology. At present, it has been applied to the route selection design of traffic roads, oil pipelines and power lines. In the traffic road route selection design, it can quickly obtain large-area three-dimensional topographic maps and image data, provide routes for designers, assist in the evaluation and comparison of multiple design schemes, and provide important support for the optimization design of routes. In the survey and design of the oil and gas industry, it can solve the problems of spatial inaccessibility and complex environment, quickly obtain high-precision terrain data and high-definition image data in the pipeline corridor, and provide basic data support for the survey and optimization of oil and gas pipeline selection. In the power route selection design, it can quickly obtain three-dimensional ground information in the route corridor and various ground object information around the route, restore the real three-dimensional scene by combining virtual technology, and assist designers in viewing the ground object and topographic information around the route from different perspectives and performing power route selection and route optimization in the three-dimensional scene. SUMMARY

[0003] The technical problem to be solved by the present application is that, in view of the technical problems existing in the prior art, the present application provides a communication line engineering survey and design method and system using unmanned aerial vehicle surveying and mapping, which has high efficiency, high automation degree and low cost.

[0004] To solve the above technical problems, the technical solution provided by the present application is as follows:

[0005] A communication line engineering survey and design method using unmanned aerial vehicle surveying and mapping, comprising the following steps:

[0006] S1, acquiring unmanned aerial vehicle image data collected by a unmanned aerial vehicle through oblique photography of a survey area; wherein the unmanned aerial vehicle image data includes visible light images and infrared images;

[0007] S2, performing preset control point piercing and aerial triangulation encryption processing on the unmanned aerial vehicle image data to generate a digital elevation model (DEM) and a digital orthographic image map (DOM) as surface model data of the surveyed area, and reconstructing a three-dimensional terrain;

[0008] S3, extracting key points in the three-dimensional terrain based on an optimized U-Net model, wherein the key points include roads and rivers;

[0009] S4, select the set of key points constituting the path in the three-dimensional terrain based on the ant colony routing selection algorithm, and then arrange and combine the multiple paths between the key points to generate multiple preliminary target routes;

[0010] S5, comprehensively evaluate the preliminary target route by the comprehensive evaluation method based on the analytic hierarchy process to obtain the optimal target route.

[0011] Preferably, the specific process of generating the digital elevation model DEM and the digital orthographic image DOM in step S2 is:

[0012] Combined with the pre-set control point information and the unmanned aerial vehicle camera calibration parameter information, spatial triangulation calculation processing is performed to finally generate the digital elevation model DEM and the digital orthographic image DOM of the entire survey area;

[0013] The generation process of the DEM is: first, the DPGr id system is used to collect feature points and lines to generate a single model DEM, then the DEM is spliced, cut to generate a single DEM, and finally the DEM is edge-connected and detected to generate the DEM of the entire survey area;

[0014] The generation process of the DOM is: using the DPGr id system to generate a single model DOM from a single model DEM, then using Ortho Mosaic to mosaic, cut to generate a single DOM, then edge-connecting, detecting and splicing the DOM, and finally using PhotoShop software to adjust the color to generate the DOM of the entire aerial photography area.

[0015] Preferably, in step S2, the specific process of reconstructing the three-dimensional terrain is:

[0016] S2.1, uniform color processing of oblique images: processing the exposure, contrast and brightness of the oblique images to ensure the modeling authenticity of the final model;

[0017] S2.2, empty three calculation on all oblique images: create an engineering for a region, and in the engineering, all photos are calculated for empty three according to the POS information;

[0018] S2.3, three-dimensional model reconstruction;

[0019] S2.4, model repair: for the model defects caused by mirror surface, blind spot and occluded point, the single body is exported and repaired in 3Dsmax under the premise of considering the workload and fidelity to ensure the three-dimensional scene restoration effect.

[0020] Preferably, in step S3, the U-Net adversarial network model is optimized by updating the generator weight based on the optimization model L1 loss function to obtain an optimized U-Net model; specifically:

[0021] The L1 loss function is called the least absolute error, which aims to minimize the sum of the absolute difference between the true value and the test value, and the formula is as follows:

[0022]

[0023] Where x i is the learned data with a length of M, y i is the corresponding label;

[0024] The L1 loss function regularization is to add an additional regularization term after the L1 loss function; the regularization term is a penalty term for the L1 loss function, which uses feature sparsification to reduce the number of weight parameters and thus reduce the complexity of the model; the regularized L1 loss function is as follows:

[0025]

[0026] Where λ is the regularization coefficient, and ||f|| p is the p-norm of the weight vector;

[0027] The conditional generative adversarial network optimizes the discriminator parameters by fixing the generator network; the generator parameters are optimized by fixing the discriminator network until the loop ends; the regularized L1 loss function and the target function of the U-Net adversarial network model can be combined to optimize the weight vector, thereby improving the robustness of the model.

[0028] Preferably, in step S3, extracting roads and rivers in the three-dimensional terrain includes road and river detection, and road and river center line extraction; wherein the road and river center line extraction problem is modeled as a discrete normalized distance label prediction problem; then, the ordered regression quantity is learned to predict the discrete normalized distance label; specifically:

[0029] In the infrared image, the normalized distance from pixel i to the nearest road or river edge is predicted; the normalized distance dn i is defined as follows:

[0030]

[0031] Where w i is the road or river width of pixel i; dn i is proportional to the probability of i on the road or river center line; at the same time, if i is located on the road or river center line, dn i is the local maximum in the direction perpendicular to the road or river; each dn K-1 is further quantized to one of K+1 intervals using a threshold {t0, t1,..., t i}; after quantization, each i is assigned a discrete normalized distance label li :

[0032]

[0033] Preferably, the road, river topological loss LT is adopted to punish the gap in the road, river detection prediction and the stray part in the road, river center line extraction prediction.

[0034] Preferably, in step S4, when the preliminary target route is generated, it is determined whether the route is optimal according to relevant factors, wherein the relevant factors include route length, engineering cost, operation time and construction difficulty;

[0035] In an alternative area, after the starting routing point and the ending routing point are specified, a mesh graph formed by routing nodes is formed for a plurality of target routing points in the area; for each line segment, the route length, the engineering cost, the operation time and the construction difficulty factor are recorded; and an optimal path is formed from each independent factor.

[0036] For the construction difficulty, the shortest path of the construction difficulty of each area and the external routing points are all connected to form a total target route; similarly, the optimal paths of other factors are formed; finally, for the target routing points in the target area, a plurality of target optimal routes are obtained, and the analytic hierarchy process is used for comprehensive evaluation, and one optimal route meeting the actual requirements is selected from the plurality of target optimal paths.

[0037] Preferably, the specific process of step S5 is as follows: firstly, a comprehensive evaluation model is established, appropriate indexes are proposed, the meanings of the indexes are accurately defined, and the index calculation method is found; secondly, the calculation method of the index weight parameter is proposed, and the index weight is determined; finally, the linear weighting method is used to evaluate the design by using the comprehensive evaluation model, and the communication line engineering design path suggestion is given.

[0038] The application further discloses a computer storage medium, which stores a computer program, and the computer program executes the steps of the method when being run by a processor.

[0039] The application further discloses a communication line engineering survey and design system using unmanned aerial vehicle surveying, comprising:

[0040] A data acquisition subsystem is configured to acquire unmanned aerial vehicle image data collected by the unmanned aerial vehicle through tilt photography in the survey area; wherein the unmanned aerial vehicle image data comprises visible light images and infrared images.

[0041] A map reconstruction subsystem is configured to perform preset control point puncturing and aerial triangulation encryption processing on the unmanned aerial vehicle image data, generate a digital elevation model (DEM) and a digital orthographic image (DOM), and reconstruct a three-dimensional terrain as the ground model data of the surveyed area.

[0042] An intelligent recognition subsystem, which extracts key points in 3D terrain based on an optimized U-Net model, including roads and rivers;

[0043] The route planning subsystem is used to select a set of key points that constitute a path in the three-dimensional terrain based on the ant colony routing algorithm, and then arrange and combine multiple paths between the key points to generate multiple preliminary target routes;

[0044] The comprehensive evaluation subsystem is used to conduct a comprehensive evaluation of the preliminary target route through a comprehensive evaluation method based on hierarchical analysis to obtain the optimal target route.

[0045] Compared with the prior art, the advantages of the present invention are:

[0046] The present invention adopts the idea of ​​system integration and reconstructs the survey and design process of communication line engineering. It transforms the current traditional on-site survey process based on designers, namely "determining the routing direction, manual on-site survey, and optimizing the line routing", into a survey and design process based on drone mapping, namely "determining the routing direction, drone mapping-assisted design, and system push of multiple preliminary routing solutions". It gives full play to the advantages of the information collection system, data analysis system and control management system, realizes the open interconnection of multiple systems, and enables the entire drone mapping data to support the entire process of line design preparation, design process and post-design completion.

[0047] The design method of the present invention has the following advantages: 1. High efficiency: UAV flights replace manual site surveys, saving time; 2. Rich data: UAV map reconstruction technology is used to quickly obtain two-dimensional and three-dimensional survey scenes; 3. High degree of automation: Deep learning technology is used to assist in survey data analysis, reducing labor costs and data processing time; 4. Low overall cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The figure is a flow chart of an embodiment of the design method of the present invention.

[0049] Figure 2 This is a schematic diagram of the division of the aerial survey area in the present invention.

[0050] Figure 3 This is a diagram of the drone aerial survey process within a single operating area in the present invention.

[0051] Figure 4 This is a schematic diagram of path selection in the present invention.

[0052] Figure 5 Schematic diagram of the evaluation of multiple design solutions for path selection in the present invention.

[0053] Figure 6Flow chart for making three-dimensional digital elevation model in the present application.

[0054] Figure 7 Mission planner ground station photographic track planning chart in the present application.

[0055] Figure 8 Single-lens aerial photographic track planning chart in the present application.

[0056] Figure 9 Basic flow chart for generating DEM and DOM in the present application.

[0057] Figure 10 Network structure chart of U-Net in the present application.

[0058] Figure 11 Process chart for training optimized road and river detection model in the present application.

[0059] Figure 12 Route chart generated by ant colony algorithm in the present application.

[0060] Figure 13 Hierarchical structure chart in the present application. DETAILED DESCRIPTION

[0061] The present application is further described below in conjunction with the accompanying drawings and specific embodiments.

[0062] As shown in the drawings, Figure 1 the communication line engineering survey and design method using unmanned aerial vehicle surveying of the embodiment of the present application comprises the steps of:

[0063] S1, acquiring unmanned aerial vehicle image data collected by the unmanned aerial vehicle through the way of oblique photography on the survey area; wherein the unmanned aerial vehicle image data comprises visible light image and infrared image;

[0064] S2, performing preset control point piercing and aerial triangulation encryption processing on the unmanned aerial vehicle image data, generating digital elevation model DEM and digital orthographic image chart DOM as the ground model data of the surveyed area, and reconstructing the three-dimensional terrain;

[0065] S3, extracting key points in the three-dimensional terrain based on the optimized U-Net model, wherein the key points include roads and rivers;

[0066] S4, selecting the set of path key points in the three-dimensional terrain based on the ant colony routing selection algorithm, and then arranging and combining multiple paths between the key points to generate multiple preliminary target routes;

[0067] S5, performing comprehensive evaluation on the preliminary target routes by the comprehensive evaluation method based on the analytic hierarchy process to obtain the optimal target route.

[0068] Specifically, in step S1, by analyzing the inspection task requirements, the configuration scheme of the unmanned aerial vehicle platform is designed, the unmanned aerial vehicle is autonomously patrolled according to the approximate routing direction of the communication line, and aerial survey image acquisition is realized.

[0069] In the process of using unmanned aerial vehicle aerial survey to assist optical cable line survey and design, first, the survey area is determined, and then the unmanned aerial vehicle aerial survey flight carries task load (camera and lens specifications) and plans flight path (flight route, flight height, and flight line overlap) according to the size of the survey area and the ground resolution accuracy requirement.

[0070] Secondly, the topography and spatial distribution of the unmanned aerial vehicle aerial survey area need to be analyzed, the planar image of the unmanned aerial vehicle aerial survey is processed by using terrain reconstruction technology, the surface model data of the surveyed area is obtained, which is used to describe the latitude, longitude, and elevation of each element in the area along the line, and the surface texture, so as to obtain the slope and topographic features.

[0071] Among them, the basic implementation process of unmanned aerial vehicle aerial survey mainly involves two steps:

[0072] (1) After determining the start and end points of the to-be-surveyed line area, the line flight path of the unmanned aerial vehicle aerial survey is preliminarily designed by using manual design and map-assisted methods, fully considering road, terrain, construction, and other conditions, and multiple independent operation areas are divided according to the endurance of the unmanned aerial vehicle. The division of a certain flight path area is shown in Figure 2 . Then, the overall area of the preliminary planning route is determined by using the online map, and the alternative aerial survey area is selected in some areas.

[0073] (2) Ground control points are arranged in each independent area for subsequent aerial triangulation process, and then the unmanned aerial vehicle follows the planned route and can adopt multiple times of collaborative aerial survey for each independent operation area in the flight path. The aerial survey adopts the tilt photography mode to photograph the topographic features of the survey area. The unmanned aerial vehicle aerial survey process in a single operation area is shown in Figure 3 .

[0074] Specifically, in step S2, the ground control points and aerial triangulation processing of the unmanned aerial vehicle aerial photograph are completed, and the aerial triangulation results are exported. The oriented unmanned aerial vehicle aerial photograph is densely matched and digitally differentiated, and the digital elevation model DEM and orthographic image DOM of the survey area terrain are generated.

[0075] Specifically, in step S3, by identifying typical terrain targets in the map, the accuracy and efficiency of the communication line survey design can be greatly improved. Since the communication line is often laid near the edge of the road and needs to avoid rivers, lakes and other terrain as much as possible, the automatic identification of road and river targets in the map can prompt and assist experts in efficient and accurate mapping during the selection and mapping of the geographic route between the key nodes of the DOM orthographic image. After unmanned aerial mapping, artificial neural networks are used to identify road and river targets in the unmanned aerial orthographic image to assist the efficient implementation of communication line survey design.

[0076] Specifically, in step S4, the routing planning rules are formulated in accordance with the optical cable line design standard specification; by comprehensively considering the cost, distance and terrain weight factors, the position selection of some key path points needs to be determined manually. Then the regional path points are calculated by ant colony optimization algorithm, etc. to realize regional path selection, which can form several alternative route design schemes. The path selection scheme is shown in Figure 4

[0077] According to the optical cable line design standard specification, the routing planning rules are formulated, combined with the identified typical terrain targets, the ant colony routing selection algorithm is used to formulate the preliminary route, and then the artificial auxiliary means is used to complete the preliminary route selection of the predetermined optical cable line.

[0078] According to the optical cable line design standard specification, the routing planning rules are determined, that is, by formulating the principles and priority order of optical cable line routing selection, combined with the identified typical terrain targets, the paths available between each key point are determined. The available paths between key points can be determined by direct connection, extension of roads, extension of existing poles and lines, etc. Determining the available paths between key points will generate multiple key points and multiple paths between key points. Ant colony routing algorithm is used to select the multiple paths between key points and to combine and rank them to formulate multiple preliminary routes, and then artificial auxiliary means is used to complete the preliminary route selection of the predetermined optical cable line.

[0079] Specifically, in step S5, the line design scheme is formulated according to the preliminary route selection result, the engineering cost estimate is carried out, and the engineering design evaluation indexes such as distance length, operation time, construction difficulty and engineering cost are comprehensively considered to comprehensively evaluate and analyze several design schemes, and the design scheme optimization is realized. Based on the multi-design scheme evaluation technology, the evaluation process of multiple line construction schemes is shown in Figure 5

[0080] ​​The application applies a system integration idea, reconstructs a communication line engineering survey design process based on an unmanned aerial vehicle surveying and mapping, and fully gives play to the advantages of an information collection system, a data analysis system and a control management system.

[0081] The design method has the following advantages: 1. high efficiency: unmanned aerial vehicle flight replaces manual point setting, time saving; 2. rich data: the unmanned aerial vehicle map reconstruction technology is adopted to quickly obtain a survey two-dimensional and three-dimensional scene; 3. high automation degree: the deep learning technology is adopted to assist survey data analysis, and manual cost and data processing time are reduced; 4. low comprehensive cost.

[0082] The application further discloses a communication line engineering survey design system adopting unmanned aerial vehicle surveying and mapping, which comprises a data collection subsystem, a map reconstruction subsystem, an intelligent identification subsystem, a line planning subsystem and a comprehensive evaluation subsystem.

[0083] The data collection subsystem is composed of an unmanned aerial vehicle and an unmanned aerial vehicle route planning software, and the unmanned aerial vehicle can automatically fly and collect data according to the route planning software.

[0084] The map reconstruction subsystem processes the original image data collected by the data collection subsystem, and generates the basic data required for construction surveying in real time and quickly, including a DOM (orthophoto map), a DEM (digital terrain map) and a three-dimensional dense point cloud. The map reconstruction subsystem provides an API for the business subsystem, and the business subsystem can obtain the above various basic data through the API.

[0085] The intelligent identification subsystem automatically analyzes the orthophoto map generated by the map reconstruction subsystem based on a deep learning technology, extracts the interest objects in the image map and converts them into a vector model to give geographical position information. The intelligent identification subsystem provides an API for the business subsystem, and the business subsystem can obtain the automatic identification result in a vector form through the API. For example, the intelligent identification subsystem can identify the road network and river features in a region.

[0086] In the line planning subsystem, a bionics ant colony algorithm (ACO) is adopted to realize the selection problem of the shortest path; automatic planning and manual assistance are combined to guarantee the implementability; the preselected line and the alternative line are combined to divide into pieces and simplify; the visual route node management and element maintenance are realized.

[0087] In the comprehensive evaluation subsystem, various quantitative indexes of the routing nodes and lines, such as length, cost, construction period and difficulty, are managed, and a multi-design scheme evaluation model based on the analytic hierarchy process is used to comprehensively evaluate the selected lines.

[0088] The design system of the present application corresponds to the above design method and also has the advantages of the above design method.

[0089] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings of the specification and specific embodiments:

[0090] S1, unmanned aerial vehicle autonomous cruise image acquisition

[0091] When the unmanned aerial vehicle carries out the communication line surveying and mapping operation, the unmanned aerial vehicle platform and the information acquisition means are important links that directly affect the task quality. Different flight platforms can meet different surveying and mapping task requirements, and vice versa. That is, for different task requirements, the unmanned aerial vehicle surveying and mapping platform should be selected and combined according to the characteristics. In order to play the advantages of different unmanned aerial vehicle platforms, the information acquisition means should be combined, and the unmanned aerial vehicle surveying and mapping platform should be reasonably selected and combined according to the surveying and mapping task requirements.

[0092] (1) Unmanned aerial vehicle platform selection

[0093] The unmanned aerial vehicle platform is a platform for flight and power output of the unmanned aerial vehicle according to its function positioning in the unmanned aerial vehicle surveying and mapping system. Whether the machine structure is a multi-rotor or a fixed wing, it is composed of a machine body, a flight control system, a power energy system, a data transmission system and the like. From the task requirements, the key indicators of the unmanned aerial vehicle platform such as the flight range, the endurance time, the flight height and the flight speed should be focused on.

[0094] Therefore, for different task requirements, the unmanned aerial vehicle platform should be selected and designed according to the characteristics, or the unmanned aerial vehicle platform should be reasonably designed and selected. According to the factors such as the planned surveying and mapping distance, the area, the time limit and the like, the type of the unmanned aerial vehicle platform is determined and the flight path is planned.

[0095] (2) Unmanned aerial vehicle oblique photography aerial topographic surveying and mapping

[0096] The basic principle of the aerial topographic mapping technology by using the tilt photography of the unmanned aerial vehicle is to use one or more cameras carried by the unmanned aerial vehicle to take pictures of the task area from multiple angles in the air, and to generate a three-dimensional digital elevation model (DEM) according to the coordinate position, camera angle, lens size and other information recorded during the shooting, the coordinate position, spatial relationship and other parameters of each pixel point of the model are matched with the actual environment, and the basic process is as shown in Figure 6 .

[0097] Due to the complexity of the regulations related to aerial photography, the time efficiency and convenience of data processing are considered, and the regulations are simplified and optimized. After many aerial survey experiments in E' Shan County of Zhumadian City, Chang'an County of Xi'an City and Jingning County of Pingliang City, it is found that only the core indicators such as camera focal length, frame size, flight relative height and shooting overlap degree need to be considered, which can improve the DEM generation speed under the premise of meeting the accuracy requirements. Among them, the camera focal length and frame size are hardware indicators, which are directly determined by the selected aerial equipment; the shooting overlap degree includes the heading overlap degree and the lateral overlap degree, the heading overlap degree refers to the percentage of the same ground image overlap degree on the adjacent images of the same flight line, and the lateral overlap degree refers to the percentage of the same ground image overlap degree of two adjacent flight lines. Among them, the flight path parameters are automatically calculated by the unmanned aerial vehicle ground control station such as Mission Planner (MP), as shown in Figure 7 .

[0098] After the aerial photography parameters are determined, the flight path planning and data collection phase can be entered, and the advantages of long range and high altitude of the fixed-wing unmanned aerial vehicle are used to optimize the operation requirements. The following methods are proposed: using a wide-angle fixed-focus lens on the unmanned aerial vehicle to improve the heading overlap degree and the lateral overlap degree, and using the vertical orthographic method to collect images, so that each pixel point can be captured from at least 9 different angles; in the flight path setting, two types of flight paths are set, the first type of flight path is the standard flight path, the main flight paths are parallel to each other, and the second type of flight path is the auxiliary flight path, which is perpendicular to the main flight path, and is used to optimize the aerial photograph coordinate accuracy, as shown in Figure 8 .

[0099] After all the images are collected, the model generation can be performed by importing the aerial survey modeling program, and the ground control points are laid out to correct the model coordinate elevation accuracy, but in the actual process, due to the complex terrain, it is difficult to lay out the ground control points. It has been proved through many experiments that at a flight height of 200 meters or more, setting a high image overlap degree and a high shooting frequency can ensure that each pixel point can be reflected in at least 9 angle images, and there is no blind area or dead angle in the terrain. Therefore, only the orthographic single-angle photographed images can be used to generate the terrain, and the generated terrain and landform features are accurate.

[0100] S2, three-dimensional terrain reconstruction based on unmanned aerial vehicle photogrammetry

[0101] Image data of the survey area is collected by using the unmanned aerial vehicle oblique photography, and then the unmanned aerial vehicle image data needs to be pre-control point pricking and aerial triangulation encryption processing, to generate digital elevation model DEM and digital orthographic image DOM as the ground model data of the surveyed area. This process is to reconstruct the three-dimensional terrain, mainly involving aerial triangulation, DEM and DOM data generation, as follows:

[0102] (1) Technical analysis of aerial triangulation, DEM and DOM data generation

[0103] 1) Analysis of aerial triangulation

[0104] The aerial triangulation is needed for the terrain reconstruction of the unmanned aerial vehicle photogrammetry, which is a measurement method for obtaining the spatial coordinates (latitude, longitude and elevation) of the encrypted points by using the field control points for encryption during the photography process. The bundle method is generally used for analyzing aerial triangulation, which takes a single image as the basic unit of adjustment calculation and is a very rigorous control point encryption method in theory, and the obtained results have the highest precision, but the calculation amount is also relatively large.

[0105] The bundle method of aerial triangulation is based on the collinearity condition equation, and its basic method is: after measuring the coordinates of each control point and the image point coordinates of the encrypted points on the image, the approximate values of the exterior orientation elements of each image and the coordinates of the encrypted points in the region are determined through the calculation of the regional network. Then, according to the collinearity condition, the error equations of the encrypted points and the control points are listed respectively, and the unified adjustment calculation of the whole region is carried out to solve the exterior orientation elements of each image and the ground coordinates of the encrypted points. The approximate values of the image exterior orientation elements and the ground point coordinates are:

[0106]

[0107] x, y are the image plane coordinates of the image point; x0, y0, f are the interior orientation elements of the image; X S , Y S , Z S are the object space coordinates of the camera station; X A , Y A , Z A are the object space coordinates of the object point; a i , b i , c i (i = 1, 2, 3) are the nine direction cosines composed of three exterior orientation angle elements of the image.

[0108] 2) Generation of DEM and DOM data

[0109] Digital Elevation Model (DEM) is a data set of plane coordinates (X, Y) and elevation (Z) of regular grid points in a certain range. It mainly describes the spatial distribution of topography of the area photographed by aerial images, and is a virtual representation of topography, which can derive contour lines, slope maps and other information. Digital Orthophoto Map (DOM) is a digital orthophoto data set generated by orthorectification, edge joining, color adjustment and mosaicking of aerial images using Digital Elevation Model (DEM).

[0110] Using unmanned aerial vehicles to take aerial photographs of the area to be surveyed in the communication line design can obtain a series of digital aerial images. Combined with the pre-set control point information and camera calibration parameter information, spatial triangulation calculation and processing are performed to finally generate DEM and DOM data sets and images of the entire surveyed area. In the process of aerial triangulation, the professional software DPGrid is used to automatically calculate the aerial triangulation data to directly obtain the grid point data set of the digital elevation model DEM. The orthophoto map DOM is generated using the data editing software Ortho Mosaic. The basic process of generating DEM and DOM is shown in Figure 9

[0111] The generation process of DEM: using DPGrid system to first collect feature points and lines to generate DEM of single model, then splice DEM, cut to generate single DEM, and finally generate DEM of the entire aerial photography area.

[0112] The generation process of DOM: using DPGrid system to generate DOM of single model based on DEM of single model, then using Ortho Mosaic to mosaic and cut to generate single DOM, then performing edge joining, detection and splicing on DOM, and finally using PhotoShop software to adjust color to generate DOM of the entire aerial photography area.

[0113] (2) Implementation of three-dimensional terrain reconstruction based on oblique photography

[0114] Using ContextCapture modeling software, the present application implements reconstruction of three-dimensional terrain based on oblique photography, and the specific implementation process is as follows:

[0115] 1) Uniform color processing of oblique images: due to the influence of weather, camera, sunlight, etc. during shooting, the exposure, contrast, light and shade of the oblique images need to be processed to ensure the modeling accuracy.

[0116] ​2) Aerial Triangulation: Triangulation is performed on all oblique images. A project is created for a region, and in the project, all photos are triangulated based on the POS information. If a photo is not fully captured, a new photo needs to be added, and a new project needs to be created and triangulation needs to be performed again. Sometimes the results of a single triangulation are not satisfactory, and secondary or tertiary triangulation is needed. After triangulation, the region can be selected for the next step.

[0117] 3) 3D Model Reconstruction: 3D model reconstruction can construct different model formats as needed. When doing terrain modeling, use orthographic images to create a project, which can generate a DSM format file. When doing scene modeling, use oblique images to create a project, which can generate an OSGB format file. 3D model reconstruction can be distributed computing with multiple nodes. To make full use of network nodes and reasonably distribute blocks, adaptive block selection can be chosen. Finally, the required terrain file and terrain model file can be generated.

[0118] 4) Manual Model Repair: For model defects caused by mirrors, blind spots, and occluded points, after singulation and export, manual repair is performed in 3Dsmax to ensure the restoration effect of the three-dimensional scene under the premise of considering workload and realism.

[0119] S3, Road and River Target Recognition Based on Artificial Neural Network

[0120] In traditional methods, multiple satellite remote sensing images are used to recognize road and river targets, but remote sensing images have the disadvantages of long update cycle, low ground resolution, and incomplete three-dimensional information. The UAV system can carry out surveying and mapping at any time according to demand, and has the advantages of comprehensive coverage of ground information, strong controllability, low cost, high flexibility, and convenient operation. UAV aerial images have comprehensive coverage of ground information and high image resolution, and image processing methods have the ability to detect roads and rivers.

[0121] Through the above analysis, the scheme for road and river target recognition is: using a UAV carrying a visible light photography pod and an infrared photography pod to collect geographic information in multiple dimensions, making up for the shortcomings of using visible light imaging alone. By using the most accurate U-Net or DeepLab artificial neural network semantic segmentation tool, roads, rivers, and other information can be detected from aerial images and infrared images respectively. Further, information fusion recognition of the two in a deep learning network is proposed to fully extract useful information and improve the success rate and stability of detection and recognition.

[0122] The main network of Unet is VGG, and its structure is as follows: Figure 10The first part of the network is called down, or it can be understood as the input image features are extracted from different levels with convolutional blocks and max-pooling downsampling. The second part of the network includes upsampling, concatenation and regular convolution operations. The CNN upsampling, i.e. expanding the dimensions of the feature map to the same dimensions as the left concatenation block. It can be seen that the gray and green arrows in the figure represent the process of linking two feature maps together. Compared with other fully convolutional segmentation networks, the main contribution of U-Net is to connect the down part high-resolution features with the up-sampling features when sampling and deepening on the network, so as to make the following convolution process better feature centralization and feature learning representation. Since upsampling is a diffusion operation, good feature centralization representation needs to be performed in the previous stage. The idea of similar hierarchical fusion is also in FPN (Feature Pyramid Network).

[0123] (1)Unmanned aerial vehicle aerial image road and river detection based on optimized U-Net model

[0124] Although the U-Net based network model has good detection effect on most Cityscape datasets, it still has some deficiencies on complex unmanned aerial vehicle aerial images, such as the detection results generated by the generator are not humanized enough, and some regions are segmented relatively fuzzy. Therefore, the U-Net adversarial network model is optimized by updating the generator weight through optimizing the model loss function, and the robustness of the model is improved.

[0125] 1) Loss function optimization model

[0126] The L1 loss function is called the least absolute error, which is to minimize the sum of the absolute difference between the true value and the test value, and the formula is as follows:

[0127]

[0128] Where x i is the length of the learned data M, and y i is the corresponding label.

[0129] L1 loss function regularization is to add an additional regularization term after L1 loss function. The regularization term is a penalty term for L1 loss function, which uses feature sparsification method to reduce the number of weight parameters, and then reduces the complexity of the model. The regularized L1 loss function is as follows:

[0130]

[0131] Where λ is the regularization coefficient, and ||f|| p is the p-norm of the weight vector.

[0132] The conditional generative adversarial network optimizes the discriminator parameters by fixing the generator network; the generator parameters are optimized by fixing the discriminator network until the end of the cycle. During the process, the flexibility of the model is limited and the sensitivity of the network parameters is reduced due to the fixed generator or discriminator network. Experiments show that the L1 distance balances the parameter vectors in the network by optimizing the generator weights. The regularized L1 distance works better. Combining the regularized L1 loss function with the objective function of the U-Net adversarial network model can optimize the weight vector and improve the robustness of the model. The U-Net adversarial network model is similar to the conditional generative adversarial network model, which gives a condition information y (y can be any custom information) to the input of the model. The condition information y is spliced with the sample data x as the input image of the generator and the discriminator, and the generator and the discriminator are optimized by the maximum and minimum cycle operation.

[0133] Since the regularized L1 loss function will produce a blurring effect, a hyperparameter θ is added to control it. When the generator can output accurate detection results for the input original image, the model will be more sensitive to changes in parameters. The selection of the hyperparameter θ is very important, and multiple experiments are needed to find the optimal value.

[0134] 2) Road and river detection model structure

[0135] The training process of the optimized model is shown in Figure 11 The optimization process of the model loss function is divided into two parts: the original road and river image is input into the generator, and the generator outputs the road and river detection result image. This image is the result image that is constantly approaching the road and river target, and the error between this result image and the manually labeled label image is compared, and the generator weight is optimized according to the error; the output result image of the generator is input into the discriminator, and since the goal is to let the discriminator output "1" (the real road and river detection result image) for the input image, the error between the output result of the discriminator and the expected value "1" is compared, and the weight of the generator is optimized. The two parts optimize the weight of the generator together, so that the generator outputs an image that is closer to the standard road and river detection result image.

[0136] The U-Net conditional generative adversarial network belongs to a deep neural network, and a large number of road and river image samples are needed for network model training. If these parameters are randomly initialized and then the model is directly applied to actual road and river extraction, the effect is difficult to satisfy people, therefore, before the model is applied to the actual situation, the model needs to be trained in advance to evaluate the model effect. At present, only a small amount of unmanned aerial vehicle aerial image can be provided, which easily leads to insufficient model training and further performance reduction. To solve this problem, the sample set is expanded through data enhancement. Data enhancement can effectively solve the problem of difficult model training through image flipping, translation transformation, changing image gray value and other methods, improve the model classification accuracy and prevent overfitting. The same network method can be easily transferred from a visible light image to an infrared image.

[0137] (2) Target detection based on multi-source information fusion

[0138] In order to improve the accuracy of road and river detection, the present application adopts visible light image and infrared image for joint detection, and realizes the complementary advantages of different data acquisition methods through the fusion of multi-source aerial image information. Specifically, the visible light and infrared image are represented as a two-dimensional image on each frame, wherein the visible light contains three channels of RGB data, and the infrared image is a complex image containing amplitude and phase information (or real and imaginary part information). The two images have similar forms at the edges, but the visible light contains rich color information, and the infrared image has the ability to distinguish the electromagnetic reflection characteristics of the target. The two information capabilities are complementary, and can be fused at the data layer and the feature layer.

[0139] The U-Net network splices the feature map of the encoder down-sampling to the up-sampling feature map of each stage decoder, thereby forming a U-shaped structure. This structure can combine the feature map information of the encoding and decoding layers to prevent target information loss. By learning from the advantages of the U-Net network, the generator network (Residual and Pyramid U-Net, RPU-Net) is constructed by combining the residual module and the global pyramid module with the U-shaped network idea, which can effectively solve the problems of tree occlusion and building interference, and realize accurate detection of roads and rivers.

[0140] The advantage of the global pyramid pooling module is that it can aggregate context information and obtain global information. Integrating the global pyramid pooling module into the U-shaped network structure can greatly improve the segmentation precision and efficiency. Adding the global pyramid pooling module to the pooling operation at the bottom of the network encoding layer and making targeted optimization for road and river extraction can well solve the situation that some objects should not be segmented but are mistakenly segmented due to similar characteristics with roads and rivers.

[0141] (3) Target extraction based on multi-source information fusion

[0142] The road and river extraction task is mainly composed of two subtasks, i.e., road and river detection and road and river centerline extraction. Since the road and river detection task is closely related to the road and river centerline extraction task, a multi-task learning framework can be introduced to detect the road and river and simultaneously extract the road and river centerline. For the road and river centerline extraction problem, existing methods either rely on regression-based methods or classification-based methods. Regression-based methods converge slowly and local solutions are unsatisfactory. Classification-based methods ignore the fact that the closer a pixel is to the centerline, the higher the degree of error classification. In order to overcome these problems, the road and river centerline extraction problem is first converted into a discrete normalized distance label prediction problem, which can be solved by training an ordered regressor. For the road and river extraction task, previous studies mostly applied pixel loss functions, such as cross-entropy loss, which are not very ideal due to the special topological features of roads and rivers, such as connectivity. Therefore, a road and river topology loss function is proposed to improve the connectivity and integrity of the extracted road and river. The road and river-topology loss function has two key features:

[0143] 1. The road and river-topology loss function combines road and river detection prediction and road and river centerline extraction prediction, and utilizes the correlation between the two subtasks to mutually improve the two subtasks.

[0144] 2. The road and river topology loss will severely punish the gaps often appearing in the road and river detection results and the false segments easily appearing in the centerline extraction results. In the present invention, the AdamW optimizer is selected to minimize the road and river topology loss.

[0145] The road and river detection task and the road and river centerline extraction task can promote each other. The road and river detection task can provide detection clues for the road and river centerline extraction task, constrain the road and river centerline, and avoid the appearance of sundries. The road and river centerline extraction task can encourage the road and river detection task to pay more attention to the key points of the road and river, thereby enhancing the road and river connectivity. In order to fully utilize this relationship, the proposed network simultaneously learns the road and river detection task and the road and river centerline extraction task under a multi-task learning scheme.

[0146] For extracting road, river centerline from raw image data, the existing research work can be roughly divided into two categories: one is based on classification method, that is, extracting road, river centerline through classification-based method, and the other is based on regression method. The features of pixels on the road, river centerline are similar to those of pixels close to the centerline, and the features of pixels far from the road, river centerline are completely different from those of pixels on the road, river centerline. However, the error caused by misclassification of adjacent pixels of road, river centerline is the same as that caused by misclassification of pixels far from road, river centerline. In order to overcome the above shortcomings, the closer the pixel is to the road, river centerline, the more tolerant it is to its classification error. This contradiction makes it difficult for the network to converge to a better result. At present, the method based on regression learns a designed function, whose return value decreases with the distance from the pixel to the centerline. However, due to anomalies (such as annotation errors), the deep network of regression is relatively unstable, and the network trained by MSE loss will not converge to a satisfactory global solution. In order to avoid the above problems, the present application uses a method based on ordinal regression to learn a discrete normalized distance label. The present application uses ordinal loss to minimize to learn the network parameters of the road, river centerline task.

[0147] In reality, roads and rivers have unique topological properties. At present, topology is usually applied before using variational and Markov random field methods.

[0148] For road, river extraction, the proposed network jointly learns road, river detection and road, river centerline extraction tasks under a multi-task learning scheme. Road, river detection branch and road, river centerline extraction branch. The encoder shared by the two branches is used for feature extraction, thereby establishing a connection between the two branches. How the network performs the road, river detection task and the road, river centerline extraction task is introduced below.

[0149] Next, the definition of the first road, river topological loss function proposed is given. Finally, how the multi-task learning framework learns the road, river detection task and the road, river centerline extraction task based on ordinal regression while using the initially proposed road, river topological loss is introduced.

[0150] 1) Road, river detection

[0151] The road, river detection task aims to detect roads and rivers from visible light, infrared images. The output of the road, river detection task is a binary image, where the pixels seen on the road, river are 1, and other pixels are 0. In fact, most pixels of the infrared image belong to non-road, river areas. Therefore, there is a label imbalance problem in the road, river detection task. In order to overcome this problem, a weighted cross-entropy loss is used, and the weight in the weighted cross-entropy loss is presented.

[0152] 2) Road, river centerline extraction

[0153] For road, river centerline extraction, the classification-based method learns a function y(·) as:

[0154]

[0155] where f i is the feature of pixel i. The regression-based method is to learn a regressor y(·) whose value monotonically decreases as i increases its distance to the centerline, the regressor y(·) is:

[0156]

[0157] where D c (i) is the metric distance from pixel i to the pixel on the centerline that is closest to i, d M is s / 2, where s is the size of the local neighborhood used to compute the feature vector f i . In this invention, ordinal regression-based method is proposed, in the following, first, the road, river centerline extraction problem is modeled as a discrete normalized distance label prediction problem; then, an ordinal regressor is learned to predict the discrete normalized distance label.

[0158] In high resolution infrared images, the distance d i from any pixel i to the nearest road, river edge can be predicted. However, the probability that pixel i lies on the road, river centerline is not proportional to the distance d i from pixel i to the nearest road, river edge. This is because the width of road, river is diverse.

[0159] Therefore, the normalized distance from pixel i to the nearest road, river edge is predicted. The normalized distance dn i is defined as follows:

[0160]

[0161] where w i is the road, river width of pixel i. In particular, dn i is proportional to the probability that i lies on the road, river centerline. Meanwhile, if i lies on the road, river centerline, then dn i is the local maximum along the direction perpendicular to the road, river direction. Thresholds {t0, t1,..., t K-1} are used to further quantize each dn i to one of the K+1 intervals. The quantized normalized distance dn iThe reason is that the direct training of the regression deep network is relatively unstable, because outliers (annotation errors) will lead to large error terms, making it difficult for the network to converge and leading to unstable predictions. After quantization, each i is assigned a discrete normalized distance label l i :

[0162]

[0163] The discrete normalized distance label prediction is predicted by the typical method based on multi-class classification. However, this will ignore the ordinal information between the discrete normalized distance labels. In this invention, ordinal regression is used to solve the discrete normalized distance label prediction problem, and the ordered loss is further modified to adapt to the road and river centerline extraction task.

[0164] 3) Road and river topology loss

[0165] At present, cross-entropy loss is widely used in various segmentation tasks such as semantic segmentation and instance segmentation. Cross-entropy loss is a pixel-level loss and is completely local, without considering the special and complex topological features of roads and rivers. This loss will equally and independently penalize the error of each pixel, regardless of the impact of the error on the geometric shape. However, since the misclassification of pixels close to the road and river centerline will cause serious topological errors such as gaps and stray parts, these pixels are more important. In order to punish the gaps in road and river detection prediction and the stray parts in road and river centerline extraction prediction, this invention adopts a new road and river topology loss LT.

[0166] Define the road and river topology loss. In order to measure the connectivity of the road and river detection prediction, the connectivity indicator is defined as:

[0167]

[0168] Similarly, the differentiable correctness measure is defined as a measure of the correctness of the road or river centerline extraction prediction:

[0169]

[0170] where T con is the connectivity indicator of the road or river detection prediction, T cor is the differentiable correctness measure, is the true label of the road or river detection, is the predicted road or river centerline probability map of I.

[0171] Since the measurement T con is susceptible to gaps in road and river detection prediction, and the measurement T coreasily affected by stray parts in road, river centerline extraction prediction. Therefore, the road, river topology metric T Road is defined as the harmonic mean between the connectivity metric T con and the differentiable correctness metric T cor :

[0172]

[0173] The road, river topology metric measures both the connectivity and correctness of the road, river extraction result. To maximize the road, river topology metric in an end-to-end manner, the road, river topology loss L T is defined as:

[0174] L T = -log(T Road )

[0175] L T is directly calculated from the original prediction and without thresholding. Therefore, K T is differentiable on the prediction and and can be integrated into the CNN. The present invention uses AdamW optimizer to minimize the road, river topology loss.

[0176] 4) Multi-task learning

[0177] In the road, river extraction framework of the present invention, the input image I is input into a shared encoder to extract features. The feature map is input into two task corresponding decoders respectively to obtain the road, river detection prediction and the road, river centerline extraction prediction. In the road, river extraction framework of the present invention, the prediction of the road, river detection task and the ground truth Y is used to calculate the weighted cross-entropy loss which can be minimized to update the parameters of the road, river detection network. The prediction of the road, river centerline extraction task and the discrete normalized distance label map L are used to calculate the ordered loss which can be minimized to update the parameters of the road, river centerline extraction network. Y, and L are used to calculate the proposed road, river topology loss which combines the prediction of the road, river detection and the prediction of the road, river centerline extraction. The proposed road, river topology loss of the present invention takes full advantage of the correlation between the two tasks and can be minimized to make the two tasks promote each other. The whole loss function is the sum of the cross-entropy loss, the ordered loss and the road, river topology loss. By minimizing the whole loss function, the parameters of the road, river detection network and the parameters of the road, river centerline extraction network can be updated simultaneously.

[0178] S4, path selection based on ant colony algorithm

[0179] Unlike the traditional routing optimization algorithm, the path of the application is composed of key points and multi-paths between key points. The selection of the path needs to determine the key points first, and then select a path between the key points. Therefore, the ant colony routing selection algorithm is improved in the application, first based on the basic ant colony routing selection algorithm to select the set of key points constituting the path, and then arrange and combine the multi-paths between the key points.

[0180] (1) Ant colony algorithm model for path selection

[0181] In the path key point selection method based on the ant colony algorithm, the direct path cost is first used as the objective function of the ant colony algorithm, and the model is as follows:

[0182]

[0183] s.t.Path={Point1,Point2,...,Point n}

[0184] cost(Point i ,Point i+1 )=Direct(Point i ,Point i+1

[0185] Wherein, cost(Path) is the total path cost, Path={Point1, Point2,..., Point n} is the key point constituting the total path, and cost(Point i , Point i+1 )=Direct(Point i , Point i+1 ) is the direct path cost between two key points.

[0186] First, starting from the starting point, the probability of selecting the next key point is as follows:

[0187]

[0188] Wherein, τ is pheromone, η is heuristic factor, and α and β are the importance of the corresponding items. The update of the pheromone τ is as follows: From the above, the pheromone is first volatilized, and then the pheromone is increased.

[0189] τ=(1-ρ)τ+∑Δτ,0≤ρ≤1

[0190]

[0191] Wherein, p is pheromone concentration volatility coefficient, d is the direct path cost, the smaller the path cost, the more pheromone increases. Sum of delta tau represents the accumulation of all the pheromone of all routes passing through the key point in this routing process.

[0192] (2) Path selection based on ant colony algorithm

[0193] In the route planning, it is necessary to determine whether the route is optimal according to some factors or considerations. The main factors considered in the communication line engineering survey and design for path selection include route length, engineering cost, operation time, and construction difficulty. If a single factor is considered, this problem belongs to a TSP problem, which can be solved by using an ant colony algorithm.

[0194] After specifying the starting routing point and the ending routing point in a selected area, a mesh graph formed by routing nodes is formed for multiple target routing points in the area. For each line segment, the route length, engineering cost, operation time, and construction difficulty are recorded. From each independent factor, an optimal path can be formed. The line graph generated by the ant colony algorithm is shown in Figure 12 .

[0195] For construction difficulty, the shortest path of the construction difficulty of each area and the external routing point are connected to form a total target line. Similarly, the optimal paths of other factors are formed. That is, for the target routing points in the target area, four target optimal paths can be obtained. Combined with the subsequent AHP method for comprehensive evaluation, one optimal path that meets the actual requirements can be selected from the four target optimal paths.

[0196] In order to better ensure the practicality of the results, the top two optimal paths for each factor in each area are generated, and the first is connected to the first, and the second is connected to the second, to finally form two optimal paths. Then, for the four factors, eight target optimal paths are finally formed through path selection based on the ant colony algorithm.

[0197] S5, Multi-design scheme evaluation oriented to engineering design index

[0198] At present, there are many evaluation indexes and evaluation quantitative analysis methods for communication network performance. Here, by referring to the evaluation process of communication network performance, it is considered that there are three problems in the evaluation of the preliminary route selection result to develop a line design scheme: first, the evaluated object is complex, how to determine the index; second, how to quantify the index and determine the weight of each index; third, how to explain and analyze the comprehensive evaluation result of the evaluated object.

[0199] To this end, a plurality of evaluation methods are compared and analyzed, it is found that the analytic hierarchy process can better solve the above problems, and the method can decompose complex problems into several factors, the hierarchical structure of each factor is clear, and the weight parameter is scientific and accurate, therefore the application proposes a multi-design scheme evaluation oriented to engineering design indexes, and a comprehensive evaluation method based on the analytic hierarchy process is used for comprehensively evaluating the preliminary route selection result.

[0200] (1) Analytic Hierarchy Process

[0201] The analytic hierarchy process, referred to as AHP, is a decision-making method that decomposes elements related to decision-making into targets, criteria, schemes and the like, and performs qualitative and quantitative analysis on this basis. The analytic hierarchy process is a system method for optimizing decision-making of targets (multi-indexes), multi-schemes, by calculating the level single ordering (weight) and total ordering through a qualitative index fuzzy quantization method, by taking a complex multi-target decision-making problem as a system, decomposing the target into multiple targets or criteria, and further decomposing into multiple levels of indexes, and then using the method of solving the characteristic vector of the judgment matrix to obtain the priority weight of each element in a level to a certain element in the previous level, and finally using the method of weighted sum to recursively merge the most important weight of each alternative scheme to the total target, and the most important weight of the optimal scheme is the largest.

[0202] The analytic hierarchy process is more suitable for target systems with hierarchical and staggered evaluation indexes, and decision-making problems whose target values are difficult to quantify. Basic principle: The analytic hierarchy process divides the problem into different component factors according to the nature of the problem and the overall goal to be achieved, and combines the factors according to the mutual influence and membership relationship between the factors in different levels to form a multi-level analysis structure model, so that the problem is ultimately reduced to the determination of the relative importance weight of the lowest level (schemes, measures, etc.) relative to the highest level (the overall goal) or the ranking of the relative advantages and disadvantages.

[0203] The calculation process is generally divided into the following steps: 1) clearly define the problem; establish a hierarchical structure model; 2) compare each element in each layer with each other to construct a judgment matrix; 3) solve the characteristic vector of the judgment matrix and test the consistency of the judgment matrix; 4) after the consistency test is passed, determine the weighted value of each layer, if the test cannot be passed, the judgment matrix needs to be adjusted; 5) according to the analysis and calculation results, consider the corresponding decision.

[0204] (2) Evaluation and Analysis Process and Comprehensive Evaluation Model

[0205] In the present application, the specific process is as follows: firstly, a comprehensive evaluation model is established, appropriate indexes are proposed, the meanings of each index are accurately defined, and the index calculation method is found; secondly, the calculation method of the weight parameter of each index is proposed, and the index weight is determined; finally, the linear weighting method is used to evaluate the design scheme by using the comprehensive evaluation model, and the communication line engineering design path suggestion is given.

[0206] By using the analytic hierarchy process, the present application proposes to establish a comprehensive evaluation model for communication line engineering design scheme, as shown in Figure 13 , which contains 3 levels in total, specifically as follows:

[0207] 1) Target layer: the target layer index is the evaluation result to be obtained. The index layer index simulated by each scheme layer is obtained after calculation to obtain the comprehensive evaluation. The result is a value between 0 and 1. The closer the value is to 1, the better the scheme, and the closer the value is to 0, the greater the necessity of scheme optimization.

[0208] 2) Index layer: the main index criterion affecting the target. It is specifically reflected in the route length, engineering cost, operation time, construction difficulty, which needs to be quantified according to the respective calculation formula.

[0209] 1. Route length: the route length has a direct impact on the engineering cost and operation time, and also affects the length of the communication line section, and further affects the design of the station. In addition, the impact of the route length on the post-maintenance cost and obstacle hidden danger is also considered.

[0210] 2. Engineering cost: each design scheme is evaluated from the perspective of engineering cost, mainly measuring the economic benefit of the entire engineering construction, and taking the engineering cost estimation result as a reference, which can be analyzed by using the requirements of the communication engineering budget preparation.

[0211] 3. Operation time: the total operation time can be obtained according to the route length of different laying modes and the average operation time per unit length. Although the operation time cannot be completely equivalent to the construction period, it directly affects the length of the construction period, and this evaluation index is particularly important in emergency construction engineering.

[0212] 4. Construction difficulty: mainly considering the safety risks and unstable factors brought by the communication line laying in special sections.

[0213] 3) Scheme layer: the scheme layer is the object to be evaluated and compared by the model, which is composed of different schemes of line design.

[0214] (3) Evaluation matrix establishment and calculation analysis. In the specific implementation process, the evaluation matrix needs to be established and calculated and analyzed:

[0215] 1) Construct a pair comparison matrix: according to the hierarchical structure diagram and evaluation requirements determined by the model, construct a pair comparison matrix A between each layer, A=(aij m×m wherein a ij >0, m is a parameter quantity; a ij represents the ratio of the influence of the i-th index of the layer -B i to the j-th index -B j of the upper layer on the index, and the pair-wise comparison matrix is a representation of all comparison results. A "1-9 scale" is used as a comparison scale, i.e. a ij The relative scale values range from 1, 2, …, 9 and their reciprocals 1, 1 / 2, …, 1 / 9.

[0216] It should be noted that the establishment of the evaluation matrix is closely related to the evaluation environment and requirements, and the four factors have different importance in different scenarios. For example, in the case of high transmission performance requirement, the routing length is more important; and for the case of tight budget, the engineering cost is more considered in the evaluation. That is, the evaluation matrix is different under different requirements. To determine the evaluation matrix, the designer needs to accurately analyze the communication line engineering requirements.

[0217] 2) Obtaining the weighting coefficient

[0218] Generally, the "sum method" with high accuracy and simplicity is used to determine the eigenvector and the maximum eigenvalue of the pair-wise comparison matrix.

[0219] First, normalize each column vector of A to get Then, sum up each row of to get Again, normalize to get which is the approximate eigenvector, i.e. the weight vector and the weighting coefficient, and finally calculate as the approximate value of the maximum eigenvalue.

[0220] 3) Consistency check, the consistency index CI is defined as follows: When CI = 0, A is a consistent matrix; the greater CI is, the more serious the inconsistency of A is. In order to determine the allowable range of the inconsistency of A, it is necessary to find the standard of the consistency index CI of A. Introduce the random consistency index RI. When CR = CI / RI < 0.1, it is considered that the inconsistency of A is within the allowable range, and the eigenvector thereof can be used as the weight vector to complete the consistency check.

[0221] The above is only a preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments, and any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the technical field, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.​

Claims

1. A communication line engineering survey and design method using a drone surveying, characterized by, The method comprises the steps of: S1, acquiring unmanned aerial vehicle image data collected by an unmanned aerial vehicle through oblique photography of a survey area; wherein the unmanned aerial vehicle image data comprises visible light images and infrared images; S2, performing preset control point piercing and aerial triangulation encryption processing on the unmanned aerial vehicle image data to generate a digital elevation model (DEM) and a digital orthographic image map (DOM) as surface model data of the surveyed area, and reconstructing a three-dimensional terrain; S3, extracting key points in the three-dimensional terrain based on an optimized U-Net model, wherein the key points comprise roads and rivers; S4, selecting a set of path key points in the three-dimensional terrain based on an ant colony routing selection algorithm, and then performing arrangement and combination on the path key points to generate a plurality of preliminary target routes; S5, performing comprehensive evaluation on the preliminary target routes by using an analytic hierarchy process-based comprehensive evaluation method to obtain an optimal target route; In step S3, the extraction of roads and rivers in the three-dimensional terrain comprises road and river detection and road and river center line extraction; wherein the road and river center line extraction problem is modeled as a discrete normalized distance label prediction problem; then, an ordered regression quantity is learned to predict the discrete normalized distance label; Specifically: In the infrared image, predict the normalized distance from pixel i to the nearest road, river edge; normalized distance is defined as follows: where is the road, river width of pixel i; is proportional to the probability of i lying on the road, river centerline; meanwhile, if i lies on the road, river centerline, is the local maximum along the direction perpendicular to the road, river direction; a threshold further quantizes each to one of K+1 intervals; after quantization, each i is assigned a discrete normalized distance label : A road and river topology loss (LT) is used to punish gaps in road and river detection prediction and stray parts in road and river center line extraction prediction; The road and river topology loss is defined as follows: in order to measure the connectivity of road and river detection prediction, a connectivity index is defined as follows: Similarly, a differentiable correctness measure is defined to measure the correctness of road or river center line extraction prediction: wherein, is a predicted connectivity indicator for road or river detection, is a measure of differentiability, is a true label for road or river detection, is a predicted road or river centerline probability map for I. Due to the measurement is susceptible to gaps in the road, river detection and prediction, while the measurement is susceptible to stray parts in the road, river centerline extraction prediction; therefore, the road, river topology measure is defined as the harmonic mean between the connectivity measure and the differentiability measure ​ The road and river topology measure measures the connectivity and correctness of the road and river extraction result simultaneously; in order to maximize the road and river topology measure in the CNN in an end-to-end manner, the road and river topology loss is defined as: 。 2.The communication line engineering surveying and designing method using the unmanned aerial vehicle surveying according to claim 1, wherein, The specific process of generating the DEM and the DOM in step S2 is as follows: Combined with the pre-set control point information and the unmanned aerial vehicle camera calibration parameter information, spatial triangulation calculation processing is performed to finally generate the DEM and the DOM of the entire survey area; The generation process of the DEM is as follows: the DPGrid system is used to first collect feature points and lines to generate a single model DEM, then the DEMs are spliced, cut to generate a single DEM, and finally the DEMs are edge-connected and detected to generate the DEM of the entire survey area; The generation process of the DOM is as follows: the DPGrid system is used to generate a single model DOM from a single model DEM, then the Ortho Mosaic is used to mosaic and cut the DOM to generate a single DOM, then the DOMs are edge-connected, detected and spliced, and finally the PhotoShop software is used for color adjustment to generate the DOM of the entire aerial survey area. 3.The communication line engineering surveying and designing method using the unmanned aerial vehicle surveying according to claim 2, wherein, In step S2, the specific process of reconstructing the three-dimensional terrain is as follows: S2.1, uniform color processing of oblique images: the exposure, contrast and brightness of the oblique images are processed to ensure the modeling authenticity of the final model; S2.2, performing aerial triangulation calculation on all oblique images: a project is created for a region, and in the project, all photos are calculated based on the POS information to perform aerial triangulation calculation; S2.3, three-dimensional model reconstruction; S2.4, repair the model: for the mirror, blind spot, shelter point caused by the model flaws, single out after taking into account the workload and fidelity under the premise of, in 3Dsmax repair, to ensure the three-dimensional scene restoration effect. 4.The communication line engineering surveying and designing method using the unmanned aerial vehicle surveying according to claim 1 or 2 or 3, characterized in that, In step S3, the generator weight pair U-Net is updated by optimizing the model L1 loss function, and the optimized U-Net model is obtained; Specifically: The L1 loss function is called the minimum absolute error, which minimizes the sum of the absolute difference between the true value and the test value, and the formula is as follows: wherein is the length M of the learned data, is the corresponding label; L1 loss function regularization is to add an additional regularization term after L1 loss function; The regularization term is a penalty term for L1 loss function, which reduces the number of weight parameters by using feature sparsification method, thereby reducing the complexity of the model; The regularized L1 loss function is as follows: where λ is a regularization coefficient, is the p-norm of the weight vector; The conditional generative adversarial network optimizes the discriminator parameters by fixing the generator network; Fix the discriminator network to optimize the generator parameters until the loop ends; The combination of the regularized L1 loss function and the objective function of the U-Net adversarial network model can optimize the weight vector, so as to improve the robustness of the model. 5.The communication line engineering surveying and designing method using unmanned aerial vehicle mapping according to claim 1 or 2 or 3, characterized in that, In step S4, when generating the initial target route, it is determined whether the route is optimal according to relevant factors, including route length, engineering cost, operation time and construction difficulty; After specifying the starting route point and the ending route point in a candidate area, a mesh graph formed by the route nodes is formed for multiple target route points in the area; For each line segment, record its route length, engineering cost, operation time and construction difficulty factor; From each independent factor, an optimal path is formed; For construction difficulty, connect the construction difficulty shortest path of each area and the external route point to form a total target route; Similarly, the optimal path of other factors is formed; Finally, for the target route points in the target area, multiple target optimal routes are obtained, which are comprehensively evaluated by using the analytic hierarchy process, and one optimal path that meets the actual demand is selected from the multiple target optimal paths. 6.The communication line engineering surveying and designing method using the unmanned aerial vehicle surveying according to claim 1 or 2 or 3, characterized in that, The specific process of step S5 is as follows: first, establish a comprehensive evaluation model, propose appropriate indexes, accurately define the meaning of each index, and find the index calculation method; second, propose a calculation method for the weight parameters of each index, and determine the index weight; finally, use the comprehensive evaluation model to evaluate the design by using linear weighting method, and give the communication line engineering design path suggestion.

7. A computer storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-6.

8. A communication line engineering survey design system using a UAV surveying, for performing the steps of the method according to any one of claims 1 to 6, characterized in that, Comprise: A data acquisition subsystem for acquiring unmanned aerial vehicle image data collected by the unmanned aerial vehicle through tilt photography in the survey area; Wherein the unmanned aerial vehicle image data includes visible light image and infrared image; A map reconstruction subsystem for performing preset control point piercing and aerial triangulation encryption processing on the unmanned aerial vehicle image data, generating a digital elevation model DEM and a digital orthographic image DOM as the ground model data of the surveyed area, and reconstructing a three-dimensional terrain; An intelligent recognition subsystem is configured to extract key points in the three-dimensional terrain, including roads and rivers, based on an optimized U-Net model. A route planning subsystem is configured to select a set of key points constituting a path in the three-dimensional terrain based on an ant colony routing selection algorithm, and to arrange and combine multiple paths between the key points to generate multiple preliminary target routes. A comprehensive evaluation subsystem is configured to comprehensively evaluate the preliminary target routes by using an analytic hierarchy process-based comprehensive evaluation method to obtain an optimal target route.

Images